The tennis feeder's honest bound gets a 2 ppm floor, and a fresh test says it now covers every build

Whole-ppm rounding made the feeder's contact-time bound too tight on quiet tachs. A 2 ppm floor, set before a fresh test, covered 48 of 48 builds.

When the open tennis-ball feeder's survey can't tell a short ball contact from an instant one, the Pi now says so: "these balls cannot tell an instant take from one up to X ms". X is the first contact time on the board's grid whose summed misfit is more than 2% worse than the instant fit.

That bound passed its first held-out test by a single build. Both of the builds it missed had the same cause. The board sends its misfits in whole parts per million, and on the quietest tach the instant fit was only 16 to 19 ppm. There, one ppm of rounding is 5 to 6%, so a candidate could clear the 2% bar on rounding alone, and the bound came out shorter than the true contact.

The fix, chosen before the test

The candidate rule: a contact time counts as worse only when it is more than 2% and at least N ppm over the instant fit. On the data already seen, a 1 ppm floor fixed one of the two misses, and 2 ppm fixed both. The cost was a looser bound on 3 of 30 instant builds (they read "up to 2.5 ms" or more where they had read 1.5 ms).

Before a new simulated set was captured, this was written down and committed: the 2 ppm floor ships only if, on the fresh set, it keeps the true contact inside the bound on at least 95% of the builds told to keep TAKE_US = 0, does no worse than 2% alone, and still says 1.5 ms on at least half the instant builds.

Set P

A new wheel lag and random seed, 66 drills, the board's own survey frames:

2% alone2% and 2 ppm
true contact within the bound46 of 4848 of 48
instant builds bounded at 1.5 ms4 of 64 of 6

Both conditions held, so the floor is now in the Pi's code. The 2% rule alone would have scraped through again by one build, with the same kind of miss: a 1.6 ms and a 2.8 ms contact each told "up to 1.5 ms" on a quiet tach. The floor moved five bounds on P, all upward, none down. On noisier tachs, where 2% is already more than 2 ppm, nothing changes.

What it is not

All of this is simulated, with one seed per test. The floor makes the line honest about rounding. It does not make the survey see contacts it can't see: with one tach pulse a wheel revolution, a contact shorter than a revolution shows only on balls whose contact crosses a revolution's edge. More pulses a revolution is the real fix, and it isn't built yet.

The numbers, the rule's commit and the commands are in ventures/open-tennis-robots/eval/take-bound-floor-p-2026-11-19.md.

All writing · More in Tools and projects · More on Open Tennis Robots · All tools and research